Example of the testing process. We add artificial AGN point sources to galaxy images and let PSFGAN recover the original galaxy.

Abstract: The study of unobscured active galactic nuclei (AGN) and quasars depends on the reliable decomposition of the light from the AGN point source and the extended host galaxy light. The problem is typically approached using parametric fitting routines using separate models for the host galaxy and the point spread function (PSF). We present a new approach using a Generative Adversarial Network (GAN) trained on galaxy images. We test the method using Sloan Digital Sky Survey (SDSS) r-band images with artificial AGN point sources added which are then removed using the GAN and with parametric methods using GALFIT. When the AGN point source PS is more than twice as bright as the host galaxy, we find that our method, PSFGAN, can recover PS and host galaxy magnitudes with smaller systematic error and a lower average scatter ( ). PSFGAN is more tolerant to poor knowledge of the PSF than parametric methods. Our tests show that PSFGAN is robust against a shift in the PSF width of if it is trained on multiple PSF’s. We demonstrate that while a matched training set does improve performance, we can still subtract point sources using a PSFGAN trained on non-astronomical images. While initial training is computationally expensive, evaluation of PSFGAN on data is more than times faster than GALFIT fitting two components. Furthermore, it is more robust and easy to use than parametric methods as it requires no input parameters.

Potential for large pipelines and high redshift

Once PSFGAN is trained it is very fast in outputting the PS subtracted image. Furthermore it has a remarkable flexibility compared to parametric methods. Overall it is well-suited for automated analyses of large datasets.

Furthermore it is a promising tool for studying AGN and their host galaxies at higher redshift where classical methods tend to break down. With increasing redshift the contrast ratio between AGN and host galaxy tends to be higher as the intrinsic emission emerges from a bluer part of the Spectral Energy Distribution (SED) where the AGN is dominant. Also the host galaxy is affected by surface brightness dimming while the PS is not. This again increases the probability of fnding high contrast systems with increasing redshift. The range of high contrasts is exactly the range where PSFGAN has demonstrated to be superior to parametric methods.

Animals vs. Galaxies

We show that PSFGAN uses its generalized knowledge about galaxy light distributions to subtract their central point sources by running it on images of cats and dogs. Also we train the same architecture on animal images and compare its ability to subtract point sources. We find that PSFGAN can better subtract point sources from galaxies if it was trained on galaxies.

Using PSFGAN

To achieve optimal results, users should train PSFGAN on their own data and use a validation set to select the best model. However, if not enough training data is available, users can apply our pretrained models out of the box. In our work we have shown that the models pretrained on SDSS data can subtract AGN point sources from Hubble images. Using galaxies with in our test sets, we achieve comparable or better performance than parametric fitting in batch mode.

Team Members


Dominic Stark

Barthelemy Launet

Kevin Schawinski

Ce Zhang

Michael Koss

Dennis Turp

Lia Sartori

Hantian Zhang

Anna Weigel